add xnli task

[hatimbr]

Adding xnli to lm-evaluation-harness based on this PR 134

[StellaAthena]

@hatimbr Thank you for the contribution! Can you run this on a couple models with public evaluations to confirm that the reported scores are what is expected?

[hatimbr]

Hi @StellaAthena, I tested it with bloom-7b1 and bloom-560m and compared it to this evaluation.

With bloom-7b1, I got:

Task	Prompt	Version	Metric	Value		Stderr
xnli_en	GPT-3 style	1	acc	0.3335	±	0.0067
xnli_en	GPT-3 style		acc_norm	0.3285	±	0.0066

The public accuracy score was 0.3333333432674408

With bloom -560m, I got:

Task	Prompt	Version	Metric	Value		Stderr
xnli_fr	MNLI crowdsource	1	acc	0.3497	±	0.0067
xnli_fr	MNLI crowdsource		acc_norm	0.3345	±	0.0067

The public accuracy score was 0.35261043906211853

Is it good enough ? Should I run more tests ?